![Page 1: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/1.jpg)
Role of perceptual similarity in learning phonological alterna4ons
James White University of O3awa
1
![Page 2: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/2.jpg)
Role of perceptual similarity in phonological learning?
• Does perceptual similarity play a role during phonological learning? – If so, what is that role?
• Part of a larger quesCon: Does phoneCc substance play a role in phonological learning? – Are there substanCve biases1?
2 1. Wilson, 2006
![Page 3: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/3.jpg)
Steriade’s P-‐map proposal
• P-‐map: a mental representaCon of the relaCve perceptual similarity between speech sounds in a given context.1 – Perhaps based on an individual’s prior perceptual experience.
• Steriade proposed that learners are biased by the P-‐map when learning phonological pa3erns. – Phonological processes assumed, a priori, to involve minimal perceptual modificaCon.
3 1. Steriade, 2001/2008
![Page 4: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/4.jpg)
Today’s talk
• Goal: InvesCgate this issue by comparing learning models both with and without a similarity bias.
• Test case: Experimental data showing that adult learners disprefer saltatory alternaCons (White, 2014, Cogni*on).
= a type of alternaCon involving not minimal change, but excessive change.
4
![Page 5: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/5.jpg)
Roadmap
1. Overview of saltatory alternaCons
2. Overview of experimental results
3. Modeling
5
![Page 6: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/6.jpg)
1. Saltatory alternaCon
6
![Page 7: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/7.jpg)
Saltatory alternaCon
• Phonological alternaCon where an intermediate sound is “leaped over”.1
• Campidanian Sardinian:2 – /p/ ⟶ [β] / V ___V – no change for /b/.
7
[pãi] ⟶ [s:u βãi] ‘the bread’ [bĩu] ⟶ [s:u bĩu] ‘the wine’
1. White, 2014 2. Bolognesi, 1998
![Page 8: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/8.jpg)
Saltatory alternaCon
• Phonological alternaCon where an intermediate sound is “leaped over”.1
• Campidanian Sardinian:2
8
p β -‐conCnuant
-‐voice +labial
+conCnuant +voice +labial
2 feature changes
1. White, 2014 2. Bolognesi, 1998
![Page 9: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/9.jpg)
Saltatory alternaCon
• Phonological alternaCon where an intermediate sound is “leaped over”.1
• Campidanian Sardinian:2
9
p b β -‐conCnuant
-‐voice +labial
-‐conCnuant +voice +labial
+conCnuant +voice +labial
2 feature changes
1 feature difference
1 feature difference
1. White, 2014 2. Bolognesi, 1998
![Page 10: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/10.jpg)
2. Experimental results
10
Learners prefer to avoid saltatory alternaCons.1
1. White, 2014
![Page 11: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/11.jpg)
ArCficial language experiments (adult learners)
11
1. Exposure phase
[kamap] [kamavi]
Also: non-‐changing fillers like [luman] [lumani]
![Page 12: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/12.jpg)
12
1. Exposure phase
2. VerificaCon phase
3. GeneralizaCon phase
[kamap] [kamavi]
[kamap] [kamapi] or [kamavi]???
[lunub] [lunubi] or [lunuvi]???
ArCficial language experiments (adult learners)
![Page 13: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/13.jpg)
45%
Crucial results to be accounted for
13
Ambiguous SaltaCon Explicit SaltaCon
Test condiCon
Control condiCon
p
v
b 70%
f
p
v
b
p
v
b p
v
b
f 16%
21%
21%
7%
![Page 14: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/14.jpg)
3. Maximum entropy learning model
14
For previous uses, see Goldwater & Johnson, 2003; Wilson, 2006; Hayes & Wilson, 2008; Hayes et al., 2009; MarCn, 2012; others. Implemented using the MaxEnt Grammar Tool (available at Bruce Hayes’s webpage).
![Page 15: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/15.jpg)
2 things the learning model should account for
1. Saltatory alternaCons exist and thus must be learnable for the child.
2. This type of alternaCon is dispreferred by learners.
– Can we model the experimental results?
15
![Page 16: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/16.jpg)
Overview of the model
16
• Set of OT-‐style constraints • Input forms (singular words) • Candidate output forms (plural words) • Constraint violaCons
Training data (observed input-‐output pairs)
Grammar (=weighted constraints)
Provide the model
Predicted probability of each output
Prior (a priori preferred weights)
(Maxent learning)
![Page 17: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/17.jpg)
Constraint set
• Feature-‐based markedness constraints, which moCvate alternaCng. – *V [–voice] V – *V [–conCnuant] V
• Correspondence constraints banning alternaCons between specific pairs of sounds.1 – E.g., *MAP(p,v) = don’t have an alternaCon between [p] and [v].
17 1. Zuraw, 2007, 2013
![Page 18: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/18.jpg)
Why *MAP constraints?
• TradiConal faithfulness constraints (classical OT) cannot generate saltaCon.1, 2 – /p/ à [v], when [b] is legal, results in a gratuitous faithfulness violaCon.
– True, even with weighted constraints.
• Straighqorward implementaCon of the similarity bias.
18 1. Lubowicz, 2002 2. Ito & Mester, 2003
![Page 19: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/19.jpg)
Training data
• Same as the training data in the experiments.
• Ambiguous SaltaCon experiment:
19
input possible outputs # observed
p v 18
p 0
f 0
b 0
input possible outputs # observed
t ð 18
t 0
θ 0
d 0
![Page 20: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/20.jpg)
Prior (= the bias)
• Biases the learning towards certain outcomes, based on a priori assumpCons. – In this case, based on perceptual similarity. – Sor bias, not an absolute restricCon.
• The prior is Gaussian. Provide: – µ = preferred weight for each constraint. – σ2 = how Cghtly constraints are held to their preferred weight.
20
![Page 21: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/21.jpg)
Measure of similarity
• How people judge the similarity of speech sounds is likely complex.1, 2, 3
• One possiblity:
21 1. Steriade, 2001/2008 2. Mielke, 2008 3. CrisCa et al., 2013
p b
f v
voicing
cont
inua
ncy
1
1
1
1
Featural similarity
![Page 22: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/22.jpg)
Perceptual similarity • I use mutual confusability as a simplified measure of perceptual similarity. – Data from published confusion experiments with adult English speakers.1
• Example:
22
p
b f
v .153
.039
.045
.013 .088
1. Wang & Bilger, 1973
Features
p b
f v
1
1
1
1
Mutual confusability
![Page 23: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/23.jpg)
Setng the prior
• Goal: go directly from confusion proporCons to prior constraint weights. (No cherry-‐picking weights).
• SoluCon: Train up a separate maxent model intended solely to generate prior weights (µ), based on confusion probabiliCes.
• IntuiCvely: Represents the listener’s perceptual experience, a computaConal version of the P-‐map.
23
![Page 24: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/24.jpg)
Prior weights
24
Constraints Substantively Biased
*V [-voice] V 0
*V [-cont] V 0
*MAP( p, f ) 1.34
( p, b ) 2.44
( p, v ) 3.65
( b, v ) 1.30
( b, f ) 1.96
( f, v ) 2.56
![Page 25: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/25.jpg)
2 other models compared
1. An4-‐alterna4on: All *MAP constraints start with the same weight. – prior weight = average of prior weights in the substanCvely biased model.
– A priori preference to avoid any alternaCon equally, regardless of similarity.
2. Unbiased: Prior weight of 0 for all constraints.
25
![Page 26: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/26.jpg)
Prior weights
26
Constraints Substantively Biased
Anti-alternation Unbiased
*V [-voice] V 0 0 0
*V [-cont] V 0 0 0
*MAP( p, f ) 1.34 2.27 0
( p, b ) 2.44 2.27 0
( p, v ) 3.65 2.27 0
( b, v ) 1.30 2.27 0
( b, f ) 1.96 2.27 0
( f, v ) 2.56 2.27 0
![Page 27: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/27.jpg)
Learning procedure
• Find the set of constraint weights that maximizes this funcCon:
• Model will provably always succeed at finding the “best” grammar by this criterion.1
27
[ ∑ log Pr (yj | xj) ] −
n
j=1
(wi − µi)2
2σi2
[ ∑ m
i=1 ]
Maximize the likelihood of the training data
with a penalty for weights that vary from the prior
1. Berger et al., 1996
![Page 28: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/28.jpg)
45%
Returning to the crucial experimental results
28
Ambiguous SaltaCon Explicit SaltaCon
Test condiCon
Control condiCon
p
v
b 70%
f
p
v
b
p
v
b p
v
b
f 16%
21%
21%
7%
P: 64%
P: 41%
P: 22%
P: 12%
P: 27%
P: 7%
![Page 29: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/29.jpg)
Model performance
29
0
20
40
60
80
100
0 20 40 60 80 100
Exp
erim
enta
l res
ults
Model predictions
r2 = .95
Recall: No manipulation of prior weights by hand!
SubstanCvely biased model
![Page 30: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/30.jpg)
Compared to other models
30
0
20
40
60
80
100
0 20 40 60 80 100
Exp
erim
enta
l res
ults
Model predictions
0
20
40
60
80
100
0 20 40 60 80 100
Exp
erim
enta
l res
ults
Model predictions
0
20
40
60
80
100
0 20 40 60 80 100
Exp
erim
enta
l res
ults
Model predictions
SubstanCvely biased AnC-‐alternaCon Unbiased
r2 = .95 r2 = .67 r2 = .25
![Page 31: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/31.jpg)
Where do the control models go wrong?
• Unbiased model: – In general, predicts too much generalizaCon – no reason to avoid posiCng new alternaCons.
• AnC-‐alternaCon model: – Unable to account for subtler variaCons between segments (e.g., [b] changed to [v] more oren than [f]).
– Ends up predicCng more generalizaCon in the control case than in the saltaCon case – opposite of the actual results!!
31
![Page 32: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/32.jpg)
Discussion
• AnC-‐alternaCon model outperforms Unbiased model. – Consistent with a general preference for avoiding alternaCons (e.g., OO-‐Faith set high by default).1, 2
• SubstanCvely biased model outperforms AnC-‐alternaCon model. – Evidence that perceptual similarity plays a role even beyond a general preference to avoid alternaCons.
• General framework for looking at the role of perceptual similarity in phonological learning.
32 1. Hayes, 2000 2. Tessier, 2012
![Page 33: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/33.jpg)
Prior as a sor bias
• The prior is crucial to the model’s success. It allows the desired learning pa3ern: – AlternaCons between dissimilar sounds are iniCally dispreferred.
– But with enough training data, the prior can be overcome – they are learnable.
• The anC-‐saltaCon effect seen in the experiments seems to fall out from a more general similarity bias.
33
![Page 34: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/34.jpg)
Evidence with infant learners?
• 12-‐month-‐olds learning potenCally saltatory alternaCon generalize to intermediate sounds.1
• 12-‐month-‐old English-‐learning infants know [d ~ ɾ], but not [t ~ ɾ], despite greater support for [t ~ ɾ] in their input.2
34 1. White & Sundara, 2012, submi3ed 2. Sundara et al., 2013 (BUCLD)
![Page 35: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/35.jpg)
Thank you!
• Acknowledgments: – For help and discussion, special thanks to Bruce Hayes, Megha Sundara, Kie Zuraw, Robert Daland, Sharon Peperkamp, Adam Albright, and audiences at UCLA, the University of O3awa, and Stony Brook University.
– Thanks to my undergraduate research assistants as well as research assistants at the UCLA Language AcquisiCon Lab.
35
![Page 36: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/36.jpg)
CreaCng the prior (details)
• Input = idenCficaCon data from confusion experiments.1
• Each *MAP constraint is weighted according to how likely its sounds are to be confused for each other.
• Output weights à Prior of main model
36
E.g.,
1. Wang & Bilger 1973
![Page 37: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/37.jpg)
Effect of different confusion data
37
![Page 38: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/38.jpg)
How the weights change during learning
38
Constraints Prior weight Post-learning weight
*V [-voice] V 0 2.45
*V [-cont] V 0 1.05
*MAP( p, f ) 1.34 1.74
( p, b ) 2.44 2.94
( p, v ) 3.65 1.96
( b, v ) 1.30 2.02
( b, f ) 1.96 2.02
( f, v ) 2.56 2.56
Explicit Salta4on experiment p à v , b à b
![Page 39: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/39.jpg)
Effect of σ
39
130
prior has almost no practical effect on the constraint weights, leaving the weights to be almost
entirely dictated by the training data. As a result, all three models converge at (virtually) the
same predictions and thus have very similar r2 values.
Figure 11. Proportion variance explained (r2) by the substantively biased model, the high faith model, and the unbiased model, according to the value of !2.
4.4.5 Effect of different confusion matrices used to derive the prior
Given that the prior weights are calculated directly from confusion probabilities, the prior
weights, and thus the model’s performance, will vary depending on which confusion data are
used as input. Implicit in this design is the prediction that real language learners will learn
slightly differently depending on their own perceptual experience and language background. This
strikes me as a reasonable assumption.
For the model predictions reported above, I added together the confusion data listed in Table
2 and Table 3 from Wang and Bilger 1973 (henceforth WB). These data were chosen for several
![Page 40: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/40.jpg)
ProducCon study
40
151
that the fit between the model predictions and the experimental results is not quite as good as in
the 2AFC task (see section 4.4.3), but the fit is still very good (r2 = .87).
Figure 12. Predictions of the substantively biased model plotted against the experimental results from the production experiment. Overall r2 = .87.
For comparison, the predictions of the unbiased model are plotted against the experimental
results in Figure 13. For the best comparison, the unbiased model was also augmented with a
*ALTERNATE constraint with the same prior values that it had in the biased model (i.e., µ of 1 and
a !2 of .0001). As the plot shows, the unbiased model is much less successful at predicting the
experimental results; as a result, the fit between the model predictions and the experimental
results is much lower (r2 = .39).
![Page 41: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/41.jpg)
PredicCng probabiliCes from the grammar
41
/kamap/ *V[-voice]V
2.45 *MAP(b, v)
2.02 *MAP(p, v)
1.96 *V[-cont]V
1.05 Total
penalty Predicted
output
kamavi *
kamapi * *
Explicit Salta4on experiment p
v
b Training:
![Page 42: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/42.jpg)
PredicCng probabiliCes from the grammar
42
/kamap/ *V[-voice]V
2.45 *MAP(b, v)
2.02 *MAP(p, v)
1.96 *V[-cont]V
1.05 Total
penalty Predicted
output
kamavi *
kamapi * *
Explicit Salta4on experiment p
v
b Training:
Input form and output candidates
![Page 43: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/43.jpg)
PredicCng probabiliCes from the grammar
43
/kamap/ *V[-voice]V
2.45 *MAP(b, v)
2.02 *MAP(p, v)
1.96 *V[-cont]V
1.05 Total
penalty Predicted
output
kamavi *
kamapi * *
Explicit Salta4on experiment p
v
b Training:
Constraint weights
![Page 44: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/44.jpg)
PredicCng probabiliCes from the grammar
44
/kamap/ *V[-voice]V
2.45 *MAP(b, v)
2.02 *MAP(p, v)
1.96 *V[-cont]V
1.05 Total
penalty Predicted
output
kamavi 1.96
kamapi 2.45 1.05
Explicit Salta4on experiment p
v
b Training:
![Page 45: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/45.jpg)
PredicCng probabiliCes from the grammar
45
/kamap/ *V[-voice]V
2.45 *MAP(b, v)
2.02 *MAP(p, v)
1.96 *V[-cont]V
1.05 Total
penalty Predicted
output
kamavi 1.96 1.96
kamapi 2.45 1.05 3.50
Explicit Salta4on experiment p
v
b Training:
e (− penalty)
∑ e (− penalty)
Sum
![Page 46: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/46.jpg)
PredicCng probabiliCes from the grammar
46
/kamap/ *V[-voice]V
2.45 *MAP(b, v)
2.02 *MAP(p, v)
1.96 *V[-cont]V
1.05 Total
penalty Predicted
output
kamavi 1.96 1.96 82 %
kamapi 2.45 1.05 3.50 18 %
Explicit Salta4on experiment p
v
b Training:
![Page 47: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/47.jpg)
PredicCng probabiliCes from the grammar
47
/kamap/ *V[-voice]V
2.45 *MAP(b, v)
2.02 *MAP(p, v)
1.96 *V[-cont]V
1.05 Total
penalty Predicted
output
kamavi 1.96 1.96 82 %
kamapi 2.45 1.05 3.50 18 %
/lunub/ *V[-voice]V
2.45 *MAP(b, v)
2.02 *MAP(p, v)
1.96 *V[-cont]V
1.05 Total
penalty Predicted
output
lunuvi 2.02 2.02 27 %
lunubi 1.05 1.05 73 %
Explicit Salta4on experiment p
v
b Training:
![Page 48: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/48.jpg)
PredicCng probabiliCes from the grammar
48
/kamap/ *V[-voice]V
2.45 *MAP(b, v)
2.02 *MAP(p, v)
1.96 *V[-cont]V
1.05 Total
penalty Predicted
output
kamavi 1.96 1.96 82 %
kamapi 2.45 1.05 3.50 18 %
/lunub/ *V[-voice]V
2.45 *MAP(b, v)
2.02 *MAP(p, v)
1.96 *V[-cont]V
1.05 Total
penalty Predicted
output
lunuvi 2.02 2.02 27 %
lunubi 1.05 1.05 73 %
Explicit Salta4on experiment p
v
b Training:
![Page 49: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/49.jpg)
Exp. 1 Results
49
0
10
20
30
40
50
60
70
80
90
100
Per
cent
cha
ngin
g pl
ural
opt
ion
chos
en
PotenCally Saltatory Control
Stops FricaCves Stops FricaCves
![Page 50: Role%of%perceptual%similarity%in%learning ...ucjtcwh/index_files/White_MOLT2014.pdf · ProducCon!study! 40 151 that the fit between the model predictions and the experimental results](https://reader033.vdocuments.us/reader033/viewer/2022042023/5e7aafcd8cb434699d5286a2/html5/thumbnails/50.jpg)
Exp. 2 Results
50
0
10
20
30
40
50
60
70
80
90
100
Per
cent
cha
ngin
g pl
ural
opt
ion
chos
en (e
rror
s)
Saltatory Control
Stops FricaCves sub-‐group sub-‐group
Stops FricaCves sub-‐group sub-‐group